Zero-crossing preservation in the low-bit-rate pseudorandomly dithered quantisation of speech signals

In earlier investigations, it has been found that, at low bit rates, the dithered quantisation of speech samples in pulse-code modulation (p.c.m.) is perceptually pleasing but is less intelligible than normal fixed-level quantised p.c.m. Also, it has been suggested that a possible cause of the reduced intelligibility of the dithered p.c.m. is the effect that the dithering has on the zero crossings of the speech signal. This paper reports on an investigation into this suggestion. A number of methods of dithered quantisation are considered in which the zero crossings of the speech signal are preserved, and the results of intelligibility tests carried out to compare the various methods of dithering are presented. The statistical analysis of the results shows that, for 3, 4 and 5 bits per sample quantisation, two of the methods of dithered quantisation are significantly more intelligible than normal, correspondingly low-bit-rate p.c.m., and the advantage is better than 1 bit per sample.